The number of AI agents being deployed across big-tech companies is way too high to count at this point. With agents supporting all domains, such as HR, finance, operations, and more, semi- to fully autonomous AI agents seem ubiquitous to any industry.
However, their usage in the semiconductor industry is not widely spoken of.
Agents Already Exist?
While not termed traditional AI agents, semiconductor design and verification processes already implement tools similar to them.
“I am sure in the software design cycle where they verify the software, you could potentially spin up agents that can launch some activities,” said Kripa Venkatachalam, VP, India technical field organisation & global customer success, Cadence Design Systems, in an interaction with AIM on the sidelines of the 5th IEEE International Women in Tech Conference (WINTECHCON2024).
Emphasising on AI agents that act autonomously, Roopashree HM, global director of EDA at Texas Instruments, explained how much of it happens through tools designed to follow set rules. Further, the tools ensure compliance and automatically fix any deviation from those rules.
“So, we already have all of those. We might not call them AI agents, but that’s also because what does an AI agent do, after all? It actually has a set of rules and it has to operate within that boundary. It has to go and make sure that you are solving the problem,” said Roopashree.
In the semiconductor industry, processes such as verification and tape-out, which are critical stages in the design and manufacturing of integrated circuits, rely heavily on advanced automation. Multimodal optimisations and the optimisation of performance, power and other parameters without human intervention are already at play.
Chip Agents Exist
Interestingly, agents for chip design are already a reality. Alpha Design AI, an AI startup, built the first AI chip agent designed to transform chip design and verification workflows for hardware engineers and semiconductor companies.
The chip agent accelerates RTL code verification, debugging, and design optimisation, integrating seamlessly into existing workflows. The startup raised $3.09 million in its pre-seed funding backed by prominent investors, including executives from semiconductor companies.
AI models are boosting chip design by improving quality and productivity, and automating tasks like code generation and debugging—an example of that is NVIDIA. Dave Salvator, director of accelerated computing products at NVIDIA, stated that their AI agents accelerate Verilog code creation and streamline design and verification using LLMs.
Furthermore, agent-based systems that can reason and take action using customised circuit design tools, besides allowing solutions from interactions with experienced designers are being developed.
Generative AI and Digital Twins
While agents for chip design are one aspect of implementing generative AI in the semiconductor industry, digital twins are another important aspect that has shown promising results.
“Digital twins are pervasive, and we are the pioneers,” said Venkatachalam, hinting at the semiconductor industry being one of the first to adopt digital twins in the form of simulators.
While it wasn’t always labelled as such, the industry has long used high-end simulators to model chip processes, even before photo-digital twins were introduced. As the cost of building chips has risen, reaching up to $5 billion for a 3nm chip, there is a growing push to create digital twins of entire fabrication lines.
Roopashree explained that this involves collaborating with equipment manufacturers and the broader semiconductor supply chain to simulate complex processes. This includes lithography and doping profiles, as well as organising data in a format that supports the creation of accurate digital twins.
Similarly, data centres are adopting digital twins for their future facilities. This convergence of digital twin technologies across sectors enhances product design, testing, and development, creating more efficient and innovative processes in industries such as automotive, industrial, and data centres.
Significant developments have already occurred in the field, with big-tech companies signing agreements to scale digital twin factories.
Samsung was preparing to launch an NVIDIA Omniverse-based Fab Digital Twin to simulate fab architecture and semiconductor manufacturing, aiming to be the first to reach smart factory Level 5. The platform will begin piloting next year, demonstrating various use cases in planning and simulation.
Similarly, Intel signed an MoU with Siemens last year to collaborate on digitisation of its wafer fabs using digital twin technology.
